Learning from other minds: an optimistic critique of reinforcement learning models of social learning

Reinforcement learning models have been productively applied to identify neural correlates of the value of social information. However, by operationalizing social information as a lean, reward-predictive cue, this literature underestimates the richness of human social learning: Humans readily go beyond action-outcome mappings and can draw flexible inferences from a single observation. We argue that computational models of social learning need minds, that is, a generative model of how others? unobservable mental states cause their observable actions. Recent advances in inferential social learning suggest that even young children learn from others by using an intuitive, generative model of other minds. Bridging developmental, Bayesian, and reinforcement learning perspectives can enrich our understanding of the neural bases of distinctively human social learning.

Automated summary

Reinforcement learning models have been used to identify neural correlates of social information value, but they underestimate the richness of human social learning. Recent advances show that even young children can learn from others by simulating their mental states. Combining developmental, Bayesian, and reinforcement learning perspectives can enhance our understanding of the neural bases of human social learning.